Multi-modality machine learning predicting Parkinson's disease.

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Precision Medicine

Background:

• Personalized medicine requires accurate disease prediction and treatment strategies.
• Machine learning (ML) and multimodal data integration are crucial for advancing predictive capabilities.
• Previous research has laid the groundwork for multimodal PD risk prediction.

Purpose of the Study:

• To develop and validate an automated ML model for improved multi-omic prediction of Parkinson's disease (PD) risk.
• To investigate key predictive features, construct disease-relevant networks, and explore drug-gene interactions.
• To make complex predictive models reproducible and accessible to the research community.

Main Methods:

• Automated ML using the GenoML package on multimodal data from the Parkinson's Progression Marker Initiative (PPMI) cohort.
• Model selection, tuning using all PPMI data, and external validation on the Parkinson's Disease Biomarker Program (PDBP) dataset.
• Network construction to identify PD-specific gene communities and analysis of feature importance (UPSIT, PRS, transcripts, SNPs).

Main Results:

• An initial ML model achieved an AUC of 89.72% for PD diagnosis.
• The tuned model validated externally on the PDBP dataset with an AUC of 85.03%.
• Combined data modalities significantly outperformed single biomarker approaches, with UPSIT and PRS being highly influential.

Conclusions:

• Multimodal data integration, powered by automated ML, significantly enhances Parkinson's disease prediction accuracy.
• The developed model is suitable for identifying at-risk populations in health registries for further monitoring and testing.
• Publicly available code and results promote reproducibility and accessibility of advanced predictive modeling in PD research.